MyDoc

RAG chat over PDF/DOCX documents. The user uploads a file, the system extracts the text, indexes it in ChromaDB and answers questions through Groq using the most relevant excerpts as context.

Made by ASCII125 — Emerson Silva GitHub · LinkedIn

---

Demo Screens

Access online demo: https://demo.mydoc.outleep.com.br

Chat com a LLM Interface principal de conversa com o modelo

Upload para análise Tela usada para enviar arquivos e iniciar a análise do chat

---

Features

• One chat per document — every conversation starts with a .pdf or .docx upload; the text is split into 4000-char chunks and stored in the vector database.
• RAG via Groq — every user message triggers a similarity search and the top 4 chunks are injected as context for the LLM.
• 10-message memory kept in Redis (rolling window) plus a pinned document summary generated when the chat is created.
• Per-visitor isolation — backend-issued HttpOnly visitor_id cookie (UUID v4); chats, messages and embeddings are scoped to the browser that owns them.
• 12-hour TTL — chats are auto-purged from SQLite, Chroma and Redis. A yellow banner at the top warns the user.
• Per-visitor rate limit — 5 uploads/min and 30 messages/min via Redis (fixed 60-second window, returns 429 with Retry-After).
• Basic CRUD — list chats, list messages (paginated), send message, delete chat (cascading purge: SQLite → Chroma → Redis).
• Retro UI — old-school OS window in soft purple/yellow/pink with beveled borders; pixelated rotating-square loader for "AI thinking" and "analyzing document" states.

Integrations

Layer	Tech	Role
API	FastAPI + Uvicorn	Routing, dependencies, middleware
LLM	Groq (llama-3.3-70b-versatile)	Replies + initial summarization
Vector store	ChromaDB (HTTP)	Similarity search
Embeddings	sentence-transformers	all-MiniLM-L6-v2 (runs locally)
Chat memory	Redis	Rolling list + pinned context
Metadata	SQLite + SQLAlchemy + Alembic	chats table indexed by visitor
Parsing	pypdf, python-docx	Text extraction
Frontend	Vue 3 + Vite	SPA with retro theme
HTTP client	Axios (withCredentials)	Sends the visitor cookie

Endpoints

All routes live under /api.

Method	Path	Description
GET	/api/health	Liveness probe
GET	/api/chats	List the visitor's chats (sweeps expired ones)
POST	/api/chats	Create chat from upload (multipart/form)
GET	/api/chats/{id}/messages	List messages (limit, offset)
POST	/api/chats/{id}/messages	Send a new message (RAG + Groq)
DELETE	/api/chats/{id}	Delete chat + embeddings + history

Interactive docs: http://localhost:8000/docs.

Layout

.
├── alembic/                       # SQLite migrations
├── compose.yaml                   # chroma + redis + api
├── Dockerfile.api                 # API image (Python 3.11 + uv)
├── package.json                   # frontend (kept at root to ease Docker)
├── pyproject.toml                 # backend
├── scripts/
│   ├── run_backend.sh
│   ├── run_frontend.sh
│   ├── run_full_stack_app.sh
│   └── docker_api_entrypoint.sh
└── src/
    ├── backend/
    │   ├── api/                   # routers, schemas, deps, middleware
    │   ├── databases/             # engine, models, repos (sqlite + redis)
    │   ├── integrations/
    │   │   ├── llm_providers/     # Groq adapter (interface + impl)
    │   │   └── vector_providers/  # Chroma adapter
    │   ├── services/              # chat_service, document_parser, rate_limiter
    │   ├── app.py                 # build_app() + lifespan
    │   ├── main.py                # uvicorn entrypoint
    │   └── setup.py               # config from .env
    └── frontend/
        ├── components/            # AppShell, ChatWindow, NewChatForm, ...
        ├── composables/           # useApi, useChats, useMessages
        ├── styles/theme.scss
        └── vite.config.js

Running

Prerequisites

• Docker + Docker Compose
• Node 18+ and uv (only if running outside Docker)
• A Groq token in GROQ_TOKEN

1. Configure .env

cp .env.example .env
# edit .env and set GROQ_TOKEN

2. Bring everything up with Docker

docker compose up -d --build
docker compose logs -f api    # follow boot + migrations

API available at http://localhost:8000.

For the frontend:

npm install
npm run dev    # http://localhost:5173

3. Without Docker

docker compose up -d chroma redis    # infra only
uv sync
mkdir -p data && uv run alembic upgrade head
./scripts/run_full_stack_app.sh

Production frontend build

npm run build              # outputs dist/frontend/
npm run preview            # serves dist/ locally

VITE_API_BASE_URL controls the API URL baked into the bundle (defaults to /api).

Configuration

Every variable is documented in .env.example. Highlights:

• GROQ_TOKEN, GROQ_MODEL
• CHROMA_HOST, CHROMA_PORT, CHROMA_COLLECTION
• REDIS_HOST, REDIS_PORT
• SQLITE_PATH
• CHUNK_SIZE (4000), MEMORY_LIMIT (10), CHAT_TTL_SECONDS (43200)
• COOKIE_NAME, COOKIE_SECURE, COOKIE_SAMESITE, COOKIE_MAX_AGE
• RATE_LIMIT_UPLOADS_PER_MIN (5), RATE_LIMIT_MESSAGES_PER_MIN (30)
• CORS_ORIGINS

Notes

• The TTL is fixed from creation — it is not refreshed by activity. Deliberate choice so the "12 hours" promise stays predictable.
• Expired-chat cleanup is lazy: it runs on GET /api/chats and on individual message calls. No cron, no worker.
• The embedding model is downloaded by sentence-transformers on first run and cached in the api_hf_cache volume.